add Lipschitz heuristic/approximation

config.lua

@@ -33,6 +33,7 @@ local common_cfg = {
     -- ^ note that this now doubles the effective trials.
     time_inputs = true, -- binary inputs of global frame count
 
+    ars_lips = false,
     adamant = false, -- run steps through AMSgrad.
 
     cap_time = 300,
@@ -50,6 +51,7 @@ local cfg = {
     epoch_top_trials = 10,
     learning_rate = 1.0,
 
+    ars_lips = true,
     deviation = 0.1,
     weight_decay = 0.004,
 

main.lua

@@ -656,11 +656,12 @@ local function learn_from_epoch()
     end
     print("best deltas:", top_delta_rewards)
 
-    local _, reward_dev = calc_mean_dev(top_rewards)
-    --print("mean, dev:", _, reward_dev)
-    if reward_dev == 0 then reward_dev = 1 end
-
-    for i, v in ipairs(top_rewards) do top_rewards[i] = v / reward_dev end
+    if not cfg.ars_lips then
+        local _, reward_dev = calc_mean_dev(top_rewards)
+        --print("mean, dev:", _, reward_dev)
+        if reward_dev == 0 then reward_dev = 1 end
+        for i, v in ipairs(top_rewards) do top_rewards[i] = v / reward_dev end
+    end
 
     -- NOTE: step no longer directly incorporates learning_rate.
     for i = 1, cfg.epoch_trials do
@@ -668,9 +669,26 @@ local function learn_from_epoch()
         local pos = top_rewards[ind + 0]
         local neg = top_rewards[ind + 1]
         local reward = pos - neg
-        local noise = trial_noise[i]
-        for j, v in ipairs(noise) do
-            step[j] = step[j] + reward * v / cfg.epoch_top_trials
+        if reward ~= 0 then
+            local noise = trial_noise[i]
+
+            if cfg.ars_lips then
+                local _, dev = calc_mean_dev(noise)
+                local c0 = neg - current_cost
+                local c1 = pos - current_cost
+                local l0 = abs(3 * c1 + c0)
+                local l1 = abs(c1 + 3 * c0)
+                local lips = max(l0, l1) / (2 * dev)
+                --reward = pos / lips - neg / lips
+                local old_reward = reward
+                reward = reward / lips
+                reward = reward / cfg.deviation -- FIXME: hack?
+                --print(("trial %i reward: %.0f -> %.5f"):format(i, old_reward, reward))
+            end
+
+            for j, v in ipairs(noise) do
+                step[j] = step[j] + reward * v / cfg.epoch_top_trials
+            end
         end
     end